Elaborating a Human Rights-Friendly Copyright Framework for Generative AI

Abstract

As works are increasingly produced by machines using artificial intelligence (AI) systems, with a result that is often difficult to distinguish from that of a human creator, the question of what should be the appropriate response of the legal system and, in particular, of the copyright system has become central. If the generator of copyright protection has traditionally been the author’s creative input, AI forces us to reassess what in the creative process is special in human creativity and where the creative input lies in AI-generated works. But it also poses more fundamental questions on what the copyright system should achieve and who/what it should protect. In particular, since many human authors will potentially face the competition of these AI machines on the market, new ways of remunerating creators will have to be imagined while making sure that the copyright system does not stand in the way of these important technological developments.

This contribution analyses the copyright issues related to so-called “generative AI” systems and reviews the arguments currently being advanced to change the copyright regime for AI-generated works. To do so, the underlying human rights framing intellectual property laws are used as the starting point from which a balanced copyright framework for generative AI could (and even should) be derived. It follows from the applicable human rights framework for copyright, but also from the anthropocentric approach of human rights, that the protection of creators and human creativity must be considered the point of reference when assessing future reforms with regard to copyright and generative AI systems. This approach establishes generative AI systems as an instrument of the human creator – and not as a substitute. It also reinforces the notion that copyright should be a tool to protect creativity and creators, not a legal mechanism to secure the amortization of economic investments in AI technology. As a consequence, it is argued that the copyrightability of AI-generated outputs should be considered with utmost care and only when AI is used as a technical tool for creators in their creation process – in other words, when they can serve a human author. At the same time, AI systems are here to stay, and their development should not be inhibited, as they can have many beneficial aspects (including for creators) if appropriately regulated.

The right to train generative AI systems via machine learning technology can be derived from the right to science and culture and freedom of (artistic) expression (Arts. 19 and 27(1) Universal Declaration of Human Rights (UDHR); Art. 15(1)(a) and (b) International Covenant on Economic, Social and Cultural Rights (ICESCR); Arts. 11 and 13 EU Charter of Fundamental Rights (EUCFR)), as AI can lead to useful advances in science and the arts; moreover, it is important for human creators to be able to use outputs produced by generative AI in their creative process. This grounding is even stronger when the training is conducted for research purposes, as the training process can then also benefit from the fundamental right-to-research justification. However, since a large quantity of copyrighted works is required for the training of generative AI systems, a remuneration obligation for these uses arises from a human rights perspective, in particular when AI systems have a commercial purpose. It follows from the right to the protection of the creator’s moral and material interests (Arts. 27(2) and 17 UDHR, 15(1)(c) ICESCR; 17(2) EUCFR, 1 Protocol No. 1, 8 European Convention on Human Rights (ECHR)) that authors must be adequately remunerated for the commercial use of their works unless there is a strong justification legitimizing the use. For this reason, it is proposed that the machine learning process using copyright-protected works to train the AI gives rise to a limitation-based remuneration right to the benefit of human creators. The article also briefly explores if and when the moral interest of creators deriving from human rights protection could justify their opposition to the use of their work for the purpose of training AI systems. It is argued that the weaker the fundamental rights claim to train the AI is, the stronger the moral rights claim could be. For example, training an AI to produce works for discriminatory or racist purposes will benefit from a weaker (if any) fundamental rights protection, but will potentially raise important moral concerns of the author of the work used for training purposes.

More generally, the article concludes that in order to secure a vibrant space for culture and creativity, (finally) cherishing and putting the Human Author at the center of the copyright system is necessary (and not only to erect fences to the benefit of copyright industries, which could be the unfortunate result of the recent first broad regulatory intervention on AI by the EU, the so-called “Artificial Intelligence Act”). In doing so, it might be possible in the future to have AI-systems that serve creators and creativity, and not the other way around.

1 Introduction

“Do you believe in the human heart? I don’t mean simply the organ, obviously. I’m speaking in the poetic sense. The human heart. Do you think there is such a thing? Something that makes each of us special and individual?” This is a question put to Klara, the narrator of Kazuo Ishiguro’s novel Klara and the Sun, who is an “artificial friend” – an artificial intelligence (AI)-operated android - that, in a not-too-distant future imagined by the author is meant to replace companions for children.¹ This philosophical question also lies at the heart of the question of the protectability by copyright of AI-generated outputs by generative AI systems: is there something in the human creative process that makes it unique and different from any output generated by a machine? As more and more works are produced by machines using AI, with a result that is often difficult to distinguish from that of a human creator, the question whether the creative input of the author – that moment of creative genius generated by the human mind, guided by intuition and inspiration – should be the generator of copyright protection has become central. It requires the identification and definition of what is an act of creativity,² but also poses more fundamental questions on what the copyright system should achieve. Does the copyright system protect authors – human authors – and remunerate them as a counterpart for their contribution to collective cultural enrichment generated by their creations, or is copyright intended as an incentive to invest in the process of cultural production, no matter how the work has been generated?³

The issue is far from purely theoretical. It was at the root of a now famous dispute in the United States between Dr. Stephen Thaler⁴ and the U.S. Copyright Office, when he wanted to register a work entitled “A Recent Entrance to Paradise” produced independently by an AI called “DABUS” of which he was the owner, user and designer. Having had his application rejected by the Copyright Office for the first time in 2019, he filed a petition for reconsideration of his application with the same Office.⁵ In its decision of 14 February 2022, the Office upheld the refusal to register the work produced by the AI, pointing out that “only a human being can be considered an “author” under US copyright law, this quality being a prerequisite for the protection and registration of the work”.⁶ Dr. Thaler challenged that rejection and filed a complaint to the Federal District Court.⁷ His motion for summary judgment seeking to register a copyright for the AI-generated work was denied on 18 August 2023, the Court confirming “that the Copyright Office acted properly in denying copyright registration for a work created absent any human involvement”.⁸ This decision could be the first episode in a long judicial journey, as it is likely that other courts will be seized in several jurisdictions with similar requests aimed at recognizing the authorship of an AI, obviously with the goal of vesting its owner (or the economic actors who designed, developed and/or financed it) with ownership of the said right.⁹ Of course, copyright regimes differ on both sides of the Atlantic, but it is worth asking how such a dispute would be resolved in Europe, as it is likely that similar questions will soon be taken to court in the EU as well. Moreover, legislators on both sides of the Atlantic are already being urged to extend copyright protection or other legal tools based on exclusivity to AI-generated output and it is likely that reform proposals will be put forward in the near future.¹⁰ When designing this new copyright framework for generative AI, legislators can find guidance in the underlying human rights norms, and the principles behind them should lead the way.¹¹

2 The Human Rights Framework for Copyright and Generative AI

Human rights are included in international and regional agreements as well as in national constitutions. They bind the legislature, as they rank very high in the hierarchy of norms, forming the roots of positive law;¹² thus, human rights have to be considered by lawmakers and judges when conceiving any legal framework applicable to a new technological, social or economic situation. Rooted in ethical principles, fundamental rights offer possibilities for a balanced development of intellectual property law. In fact, the rise of the use of human rights and fundamental rights in shaping and using intellectual property norms has led in the last 20 years or so to a “constitutionalization” of intellectual property law,¹³ helping to design most recent evolutions in the digital environment under the heading of what is increasingly understood as “digital constitutionalism”.¹⁴ Despite being improvable,¹⁵ the constitutionalization of IP law has brought major advances for a balanced conceptualization and application of IP law and still can offer a useful framework for the major future developments in the field, such as for example platform regulation, online access to information and research, or artificial intelligence.¹⁶ Moreover, it offers a transparent and workable methodology for mitigating conflicts of rights and secures a perspective on innovation law rooted in the core values of the legal systems,¹⁷ a field too often subject to influence from sectorial interests. It is argued here that human rights can (even must) serve as an effective framework to define an ethical copyright regime applicable to artificial intelligence (AI), in particular generative AI.

In general, copyright protection can be derived from provisions protecting the right to science and culture,¹⁸ the right to freedom of (artistic) expression,¹⁹ the right to protection of the author’s moral and material interests,²⁰ and the right to (intellectual) property.²¹ These human rights provisions have in common an anthropocentric approach, the human author being at the center of the protection.²² Consequently, a human rights perspective mandates that the protection of creators and human creativity is considered as the point of reference when assessing future reforms concerning copyright and generative AI. With regards to the right to intellectual property under Art. 17(2) EU Charter of Fundamental Rights (EUCFR), it is to be noted that the wording of the provision as such makes no explicit reference to the human creator.²³ It could thus be argued that the scope of protection also includes beneficiaries other than the creator himself. However, Art. 17(2) EUCFR must be interpreted in the light of other international human rights provisions which clearly centralize the protection of the human creator. Further, the concept of the social function of intellectual property law must be taken into account when interpreting this provision. It follows that copyright protection should be granted to the extent it serves the public interest,²⁴ human creativity rather than investment being the triggering factor for protection, which leads also to an author-centric approach.²⁵ In contrast, the afore-mentioned constitutional rights do not grant protection for machines or for those who operate them. Generative AI as such does not enjoy human rights protection; thus, from a human rights perspective, the protectability of its results will be dependent on the degree of the human creator’s involvement.

As outputs produced by generative AI can be beneficial to human creators in their creative process, the training of AI systems is essential to enable human beings to explore new avenues of artistic expression that are still unknown. It should be made clear that freedom of (artistic) expression concerns exclusively human beings, considering that, at least under the current state of the law, AI does not enjoy the afore-mentioned constitutional rights, in particular no right to free expression.²⁶ “This implies that the interest in the flourishing of the generative AI industry remains instrumental to the end objective of increasing human artistic freedom of expression”.²⁷

Moreover, the outputs produced by generative AI can lead without any doubt to useful advances in science and the arts that benefit society at large and hence can fall under the protection of the right to science and culture.²⁸ Accordingly, a right to train AI systems for artistic and scientific purposes can be derived as a principle from the underlying human rights framework. The human rights grounding is even stronger when the training is conducted for scientific and research purposes, as the training process can benefit additionally from the fundamental right-to-research justification.²⁹

In order to train these generative AI systems, a large amount of copyrighted works is required. It follows from the right to the protection of the creator's material rights under Arts. 27(2) Universal Declaration of Human Rights (UDHR), 15(1)(c) International Covenant on Economic, Social and Cultural Rights (ICESCR), and the right to (intellectual) property, under Art. 17 UDHR, Art. 17(2) EUCFR, 1 Protocol No. 1 European Convention on Human Rights (ECHR),³⁰ that the author must be fairly renumerated in the event of the commercial use of his work in the absence of justifications to do so that derive from competing human rights.³¹ As argued previously, “a conceptualization of copyright within the constitutional right to science and culture, to freedom of expression and artistic creativity and even within a socially rooted property clause […] should lead to the recognition of an overarching and unwritten right for creators to be remunerated for the commercial exploitation of their work. This right should be understood as a fundamental and binding principle of copyright law deriving from fundamental rights and from copyright’s social function, and would thus have to be recognized by legislators through the implementation of appropriate mandatory copyright contract rules or statutory remunerations rights; it could also be used by the judiciary as an interpretation tool of existing rules to redress unfair remuneration situations for creators, or, in their absence, even beyond as a general principle of law”.³² In contrast, the mere amortization of economic investment in AI technology can under no circumstances be derived from a fundamental rights perspective.

Beyond the protection of the author´s material interests, the human rights framework also confers the protection of moral interests, under Arts. 27(2) UDHR, 15(1)(c) ICESCR.³³ More generally, moral rights can emanate from the need to protect the authors´ dignity and personality³⁴ – values that underly various fundamental rights in the human rights treaties. The protection of the author´s moral interests can be allocated to the right to privacy and personal integrity or to the so-called “negative” aspect of the right to freedom of expression.³⁵ In the tension between copyright and generative AI training, the moral right of the author could justify a right to oppose the use of his work in certain (limited) circumstances. In fact, the weaker the fundamental rights claim to train AI is, the stronger the moral rights claim can be. For example, training an AI to produce works for discriminatory or racist purposes will benefit from a weaker (if any) fundamental rights protection, but will potentially raise important moral concerns of the author of the work used for training purposes.

In short, these general principles deriving from the human rights framework, notably the centrality of the human creator, should inform the copyright reforms with regards to generative AI systems.

3 The Protectability of AI-Generated Works

Before considering the policy question of whether (or not) it is desirable to protect works generated by AI, it is interesting to analyze if de lege lata an AI-generated work can already meet the conditions for protection.

In order to be protected, it is obvious that the work generated by an AI must meet the conditions set forth by copyright law. Many (excellent) writings have recently been devoted to this issue, so that a detailed analysis will not be conducted here.³⁶ Let it just be recalled that according to copyright law, the condition is twofold: first, there must be a protectable “work” and second, this work must be “original”.³⁷ What is a “work” is generally not defined by the legislator at national, EU or international level; however, it is commonly accepted that human intervention is necessary for a work to be protected under copyright law. According to some authors, this obviously stems from copyright law’s emphasis on the creator, whose creative act is the generator of the protection granted.³⁸ In fact, the notion of originality, defined by the Court of Justice of the EU as the “author’s own intellectual creation”,³⁹ implies creative choices.⁴⁰ This inevitably necessitates the ability to make conscious decisions, thus requiring a “mind”, a “heart” or a “soul”, to use Kazuo Ishiguro’s poetic imagery quoted above.⁴¹ In a less poetic way, this is exactly what the U.S. Copyright Office recalled when it held that for a work to be protectable, it must emanate from a human being.⁴² The Office had already decided in a similar way when specifying that a photo taken by a monkey cannot be registered as a work.⁴³ Reference is made here to the famous self-portrait taken by a monkey, the “Monkey Selfie”, which had given rise to a dispute over copyright ownership of the photo between the photographer who owned the camera used by the said monkey and the association PETA (People for the Ethical Treatment of Animals).⁴⁴ Therefore, the question: “Is a work produced exclusively by an AI protected by copyright?” can be answered negatively based on current understanding of copyright law.⁴⁵

That said, it would be wrong to stop there. As a study conducted for the European Commission has pointed out, although a work solely created by an AI cannot be protected by copyright, works for which the creators use the AI as a tool in the production process can be, provided of course that the conditions of originality and the existence of creative choices are met.⁴⁶ With regard to copyright principles, this is nothing new, since it is accepted that an author can use a computer program or a machine for the realization of his/her creation from the moment he/she makes creative choices. This does not exclude an element of randomness in the result when it is consciously included in the creation process. The photographer presses the button, but it is the camera (and its various functions, often automatic, such as filters) that takes the picture.⁴⁷ Sometimes, moreover, it is the creator’s choices alone that will be at the origin of the protection granted. Contemporary art provides numerous illustrations of this. When Duchamps “chooses” a commercially purchased urinal for display, it is the choice, not the making, that constitutes the creative act at the origin of the copyright. In the context of databases, it is the creative choices in the arrangement, selection and presentation of the database that will trigger copyright protection.

To return to AI-generated “works”, human creative intervention can occur at different stages of their production: Hugenholtz et al. proposes to distinguish the “conception” (design and specifications), the “execution” (producing draft version) and the “redaction” (selecting, editing, refinement, finalization).⁴⁸ The latter corresponds to the stage of finalization of the work, which will see the AI-generated result reworked, refined, edited and modified by human intervention (which, if it is creative and meets the conditions of copyright, can lead to copyright protection).⁴⁹ As the condition of originality is usually quite easily met, judges often finding creativity in productions with a very low personal touch, the exclusion of protection for AI-generated works could be circumvented quite easily as long as there is creative human intervention in the final result.⁵⁰ For these reasons, several scholars have considered that there is no need for legislative intervention at the EU or US level because the current copyright framework already allows for protection of most AI-generated works.⁵¹

Thus, the question arising from a de lege lata perspective is how much human creative intervention on the AI-generated output is necessary to trigger copyright protection. There is certainly a grey zone here that will lead to difficult case-by-case decisions.⁵² AI technology is already frequently used by artists in their creative process, and AI-based artistic practices are likely to increase in the future.⁵³ As an illustration, the Museum of Modern Art (MoMA) has recently acquired Refik Anadol’s artwork “Unsupervised – Machine Hallucinations”, marking the first time the institution has ever added a work created using mainly AI technology to its collection.⁵⁴ According to the digital artist, he has utilized machine intelligence as a “collaborator of human consciousness […] to unfold unrecognized layers of our external realities.”⁵⁵ In order to generate the piece, Anadol and his team trained the AI with vast amounts of data (including copyright protected works) from digital visual archives of the MOMA and other publicly available sources, then output a series of ever-shifting audiovisuals that reinterpret, alter, and riff on those original works.⁵⁶

The “copyrightability” of this particular work has not (yet) been discussed or questioned; however it can be expected that such challenges will be made in the future. It is thus to be hoped that in order to distinguish the protected from the unprotected output, the human creativity factor will be decisive, and protection granted only when the human creative input is predominant in comparison to that by the machines in the end result: in short, where it can be clearly demonstrated that the machine was strictly used as a tool in a human creative process,⁵⁷ that the end result was “AI assisted” and not “AI-generated”, the burden of proof lying on the party claiming copyright protection.

Putting human creativity at the center of the copyrightability question does not, however, answer the question from a policy perspective if this should be a desirable result.⁵⁸ In fact, some authors have argued that it would be necessary to overcome the hurdle of human intervention to protect AI-generated creations.⁵⁹ The discussion is admittedly not entirely new and recalls the “copyright without author”⁶⁰-debates held at the time when copyright protection for computer programs was introduced.⁶¹ According to others, a more cautious approach needs to be taken and access to copyright protection strictly limited, at least until the desirability of an extension is clearly established.⁶² This question is obviously most complex, and the answer might depend, inter alia, on the evolution of the technology and its future uses. However, it is worth briefly recalling some of the arguments made by each side, to the extent that an evolution of copyright law is now often required.

4 The Desirability of Granting Copyright Protection to AI-Generated Works

As Professors Vivant and Bruguière have written, “de lege ferenda, the discussion is open”.⁶³ However, the decisions to be taken are so fundamental for the future of our society that the conversation cannot be had with lawyers alone; any reform would also have to rely on independent and serious impact assessments to establish its benefit. Too often in the past, legislative interventions were conducted under the influence of sectorial interests, without taking into account the potential consequences of such interventions on the innovation ecosystem or their implications on society.⁶⁴ A good example is the damage caused at the time by the proposed Directive on computer-implemented inventions, for which the European Commission ignored the vast majority of economic studies on the issue (including those commissioned by the Commission itself), which had pointed out the potential negative effects in terms of innovation that the broad patenting of computer programs might cause, and urged the legislator to be cautious.⁶⁵ The result: the Commission went ahead, leading to the most radical rejection by the European Parliament of an intellectual property legislation in its history; in the meantime, the legal uncertainty on the patentability of computer programs remained, given the unclear judicial practice of the European Patent Office on the issue, which was also detrimental to many economic players in Europe.⁶⁶

Returning to AI, studies on the potential effects on creativity of opening up copyright protection for AI-generated creations are currently lacking.⁶⁷ If AI creations were easily protected, this would mean that countless creations would potentially be granted copyright protection, since an AI has almost unlimited production capabilities and can generate an enormous quantity of new works in record time at low cost.⁶⁸ This obviously raises the question of the creative spaces remaining for human creators and the potentially very deterrent effect on future creation that such a development would have.

Moreover, a potential change in copyright law needs to be evaluated in the light of the function that we want copyright to fulfill.⁶⁹ An AI does not need to be incentivized to produce, it just implements what it has been asked to perform, and this as often as it is asked to. It does not need a break to eat, sleep, or to search for inspiration. In fact, those who emphasize the need to protect AI-generated works, whether by copyright⁷⁰ or related rights,⁷¹ do not highlight the need to reward creators but rather the objective of encouraging investment in the field of AI. This well-known mantra has become an almost Pavlovian reflex of investors in an economy now based mostly on intangible assets: “I have invested, so I must benefit from an intellectual property right”. The issue is thus closely related to the role of investment protection within intellectual property law. As Reto Hilty et al. have rightly underlined, “potential protection regimes for AI – if ever required – would not be looking at creators or inventors, but at investors”, concluding that “most AI applications lack a theoretical justification for creating exclusive rights. If this fact is ignored, such legislation could ultimately lead to dysfunctional effects that have negative impacts on social welfare”.⁷²

The quasi-systematic push for IP protection is (unfortunately) not a new issue either. As has often been underlined, intellectual property rights tend to progressively shift to become an investment-protection mechanisms.⁷³ The multiplication of neighboring and sui generis rights is surely a consequence of this trend, as is the progressive extension of the scope of protection.⁷⁴ However, this multiplication of intellectual property rights, sometimes on the same object, has not remained without consequences and has produced a “legal hamburger in which several layers of rights overlap”.⁷⁵ These layers consist of legal hurdles to be overcome in the form of multiple authorizations to use, all resulting in potential limitations for those who want to create (not to mention the uncertainty that often surrounds the ownership of rights). To take only one example, the sui generis protection for database producers: subsequent evaluations by the European Commission have failed to determine whether the introduction of this right has increased the production of databases within the EU.⁷⁶ Knowing that once an intellectual property right is in place it is almost impossible to repeal it,⁷⁷ past experiences in the EU should lead to a cautious approach before protecting AI-generated productions by copyright and granting rights to investors.⁷⁸ On the contrary, the trend toward overprotection of intellectual property rights should prompt prior reflection on the justifications for granting them.⁷⁹ If IP rights aim at encouraging creation and protecting creators, creativity – not investment – should be the triggering factor for protection.⁸⁰ On the other hand, if the goal is to create mechanisms for investment protection, then the system cannot continue to be based on the generous protection justified by the protection of the human creator, and would need reform.⁸¹

To sum up: taking the binding human rights framework as guidance for copyright protection leads to an overall cautious approach when protecting AI-generated works by copyright and/or granting rights to investors. AI-generated output “as such” lacks the traditional protectability that human-created works enjoy. An analogy can be drawn with Art. 52 of the European Patent Convention, which specifies certain items that are not regarded as inventions. One of these exclusions concerns computer programs. In a similar vein, it could be decided that AI-generated content, in its raw, unaltered form, may not be considered as a copyrighted work. This exclusion would not prevent the combination of AI and creative human input, but would set a clear signal from the legislator in particular to the judiciary in order to avoid the unwritten condition of human authorship being overruled in the future. It would also induce, in the spirit of the exclusion of patentability of computer programs “as such”, a higher level of argumentation to show that the human creative input was determinant and predominant in the final result. Only when AI serves as a tool of human creativity could it result in a copyrighted work. In such cases, the unique blend of AI's capabilities and human ingenuity can produce content that is eligible for copyright protection.

However, when determining the scope of protectability, it will be necessary to examine very closely whether these extensions of exclusive control provide sufficient spaces of freedom to guarantee follow-on creations and future creativity. If one wants to incentivize the development of AI-generated creations, it would probably be more efficient to further widen the scope of certain exceptions to copyright and database rights, such as the exceptions allowing text and data mining.

This leads us to the third crucial question with regard to AI-created works: their legality with regard to copyright law. In short, after discussing the protectability of the output of the machine, the question of the legality of the input necessary for the machine to be able to generate new works needs to be addressed.

5 The Legality of AI-Generated Content Trained on Copyright-Protected Works

AI systems do not produce from scratch. They have to be trained through the process of machine learning with existing data and works.⁸² As Martin Senftleben pointedly describes it, “generative AI systems are only capable of mimicking human creativity because human works have been used as training material. On the basis of existing literary and artistic creations that serve as input data, machine-learning algorithms are able to recognize patterns and similarities. Following this deductive method, a generative AI system learns how to produce novel literary and artistic output by imitating the style of human works. The machine-learning algorithm enables the generative AI system to generate literary and artistic content on its own – based on the computational analysis of human works that served as training material”.⁸³ In short, to produce a work “in the style of Picasso”, the machine will need to be trained on a high number of existing works by Picasso. The better (and more comprehensive) the training, the better the results obtained.⁸⁴ This of course begs the question of how this training can be evaluated from a copyright perspective, in particular when the algorithm is “fed” with copyright protected works.⁸⁵ The answer is not easy, as so far no legislation has been passed to deal exactly with this situation. Therefore, one is left with the general principles of copyright law and with the existing rules that could apply to the situation.

Before examining the European framework, it is worth having a look at how other jurisdictions might handle the tension between copyright and AI learning. In the US, it is generally considered⁸⁶ that text and data mining (TDM), which is at the core of the machine learning proc ess, falls under the fair use exception, based on the application by analogy of Authors Guild v. Google⁸⁷ and Authors Guild v. HathiTrust.⁸⁸ It has to be mentioned however that these decisions were issued before generative AI systems had bloomed. It cannot be excluded that the wide range of possible applications on a scalable level could lead US judges to a different conclusion when applying the fair-use test under US copyright law. The fourth factor⁸⁹ of the fair-use analysis, which focuses on the effects on the market, seems to be the problematic point in the AI context, as direct competition that AI-generated works can have with those created by physical persons may lead to chilling effects on human creativity and creators’ earnings.⁹⁰ In the case of Authors Guild v. Google, the plaintiffs made three arguments why Google´s service should not be considered fair use. First, they contended that Google’s service to provide digital copies of entire books, allowing users, through the snippet function, to read parts of the book, provides a substitute for the plaintiffs’ work, which would negate the possibility of a fair use finding. Second, the Authors Guild alleged that Google had infringed their derivative rights in search functions, depriving the plaintiffs of revenues or other benefits they would gain from licensed search markets. Third, Google’s storage of digital copies exposed the plaintiffs to the risk that hackers would make their books freely (or cheaply) available on the Internet, destroying the value of their copyrights. However, the Court rejected all of these arguments, the main reason why several scholars have argued that the use of copyrighted works to train the AI is likely to be considered fair use under US law.⁹¹ Nevertheless, scholars are closely monitoring several lawsuits against AI system producers in the U.S., as rightholders are claiming that these uses are unfair and therefore are not covered by the fair use exception of US copyright law.⁹² There are initial indications that the rightholder lawsuits will not be successful. For example, a US district court in a recent decision, Andersen v. Stability AI Ltd, dismissed most of the claims that images generated by the AI systems based on text prompts violated the plaintiffs’ copyrights.⁹³ The various lawsuits pursued in the US will soon bring first elements about the legality of the use of works for ML purposes. It will be interesting to see if the upcoming US cases will be influenced by the results in the EU or if the US will focus on a completely different approach.

An illustrative instance of a maximalist approach can be seen in the recent legislative proposal unveiled in France on 12 September 2023.⁹⁴ This proposal suggests placing the machine learning process under the exclusive control of rightholders whose works are used in the ML process. Furthermore, it advocates for the attribution of authorship of the AI-generated output to all the authors whose works have been used in the machine learning process. Additionally, it mandates labeling the resulting output as “AI-generated work” and listing the names of all the authors whose works were utilized in the training process. Implementing such a measure would necessitate acknowledging a substantial number – potentially thousands or even millions – of authors, given the vast datasets that generative AI commonly trains on. In addition to being impracticable, this all-encompassing solution would likely have very detrimental consequences for the advancement of AI systems. It would surely lead to making any jurisdiction adopting such a solution very unattractive for AI innovators, in particularly start-ups who do not have the capacity to clear all the rights for the works used in the training process.

An AI-guideline proposal by the Japanese government from early June 2023 is heading in the completely opposite direction. The Japanese government released some of the world’s first legal guidelines around generative artificial intelligence imagery.⁹⁵ Although not implementing a new statutory regulation, the guidelines affirm that machine learning engineers are allowed to use any data they can find, regardless of their copyright protection.⁹⁶ Japan had already implemented one of the most liberal laws in 2018, allowing the free use of copyrighted works for training machine learning models as long as the purpose “is not for enjoying or causing another person to enjoy the ideas or emotions expressed in such work.”⁹⁷ This new guideline is based on a much broader understanding of this law, allowing the use of any data “regardless of whether it is for non-profit or commercial purposes, whether it is an act other than reproduction, or whether it is content obtained from illegal sites or otherwise.”⁹⁸

In the European Union, when examining the current legal provisions, since AI systems learn from datasets using the technique of text and data mining (TDM), the most obvious provisions to be scrutinized in this respect are the newly introduced limitations and exceptions for text and data mining-purposes in the Directive of 17 April 2019 on Copyright in the Digital Single Market (CDSM Directive). Article 3 introduces an exception for text and data mining for scientific research which solely benefits research organizations and cultural heritage institutions while Art. 4, introduced later in the elaboration process of the Directive, is not restricted to specific institutions and therefore could be relevant in the context of AI, as these systems are usually operated by private commercial companies not covered by Art. 3. According to Art. 4(1), “Member States shall provide for an exception of limitation […] for reproductions and extractions of lawfully accessible works and other subject matter for the purposes of text and data mining”. Interestingly, according to Art. 4(2), “reproductions and extractions made pursuant to paragraph 1 may be retained for as long as is necessary for the purposes of text and data mining”, which could be of relevance to solving any possible question of the storage of protected works by the AI in the learning process. However, the usefulness of this provision might be rather limited by the third paragraph of Art. 4, which subjects the application of the exception to the fact that the use of works and other subject matters “has not been expressly reserved by their rightholders in an appropriate manner, such as machine-readable means in the case of content made publicly available online”. In short, rightholders can “opt out” of the exception, which potentially can make the provision rather ineffective if a high number of rightholders do so.⁹⁹ A lot has been written on the lack of ambition of the provisions on text and data mining, and it is not the place here to discuss them at length.¹⁰⁰ These exceptions have faced strong criticism for taking insufficient account of the importance of text and data mining in many sectors, in particular with regard to the development of AI activities in the EU,¹⁰¹ while other regions of the world have more flexible approaches in their system of existing limitations.¹⁰²

In any case, what needs to be emphasized here is that the TDM exceptions were not designed to cover machine learning by generative AI systems, meaning AI systems that can produce new works based on the learning from other existing works.¹⁰³ This emerges very clearly from the preparatory works of the Directive, where the issue of generative AI and TDM exceptions is absent and has obviously been overlooked. It is thus rather surprising that the Commission considers that the existing legislative framework is perfectly fit for the purpose and that “creation of art works by AI does not deserve a specific legislative intervention”, since the TDM exceptions with their possibility of opt-out apply, “providing balance between the protection of rightholders including artists and the facilitation of TDM, including by AI developers”.¹⁰⁴ The AI Act in its latest version adopted by the European Parliament on 13 March 2024 seems to echo that position when imposing obligations on providers of general-purpose AI models to “put in place a policy to comply with Union copyright law, and in particular to identify and comply with, including through state-of-the-art technologies, a reservation of rights expressed pursuant to Article 4(3) of Directive (EU) 2019/790” (emphasis added).¹⁰⁵ This regulation further introduces transparency obligations for AI developers with regard to the works used in the training process.¹⁰⁶ The vibrant public debate generated by the use of existing works by AI systems for the purpose of generating new ones together with the tremendous economic and societal impacts that these technologies will have on the lives of all citizens seem to indicate that the discussion is not over. In fact, these developments call for another future transparent legislative intervention, as it is hard to imagine that these crucial questions could be settled by legislative provisions that were not designed for this purpose. As we have seen, in various jurisdictions, lawsuits have been introduced to challenge the legality of the use of existing works to train AI systems, and the first decisions are expected soon, which could also affect the future legal framework for generative AI at global level.¹⁰⁷ A certain degree of international harmonization will then be needed in order to avoid AI developers being excessively penalized in the EU context, in particular if the uses of copyrighted works for training purposes should be considered in certain jurisdictions to fall under the fair use provision or other exceptions and limitations to copyright law, thus opening the way for a possible international legal intervention in fora such as WIPO and/or the WTO. In any case, such crucial developments should lead to informed legislative interventions following an open, evidence-based transdisciplinary debate on the economic and societal merits of various legal interventions.¹⁰⁸

Therefore, it is necessary to reflect on possible solutions in this context, as it seems evident that the existing legal framework is not easily adaptable to deal with such an important societal question as the use of existing protected works by AI systems. Fundamental philosophical questions are at stake, in particular what should be the place in the future of creators in a world where works can be created quickly and at rather low costs by AI systems. As the AI is trained with already existing works, how to then incentivize break-through creativity, new trends, new genres, disruptive art that break with the existing?

6 Towards a New Limitation-Based Remuneration Right to the Benefit of Creators for Machine Learning of Generative AI

It is surely not the place here to propose a final “ready-to-go” solution to the issue. A future regulation of AI, as we have already stated above, needs to build on a societal consensus which requires a pluridisciplinary assessment to reach informed consent. Some preliminary thoughts can however already be shared. First, applying Art. 4 of the Directive on copyright and related rights in the Digital Single Market (CDSM) and its “opt out”-mechanism to generative AI is not a satisfying solution if we do not want to inhibit the development of this technology and thus make Europe totally unattractive for AI developers.¹⁰⁹ Not only does the provision involve a lot of uncertainties¹¹⁰ (when exactly is a content online “lawfully available” to use? How exactly to exercise the opt-out, or, put differently, what is “an appropriate reservation by machine readable means”? And who should be able to decide about this, the author or its derivative rightholder?), it also seems obvious that the opt-out will be used as a new bargaining power to license the use of existing works for training purposes.¹¹¹ Given that generative AI systems are trained on an immense number of works if they are to function efficiently, getting all the relevant authorizations would quickly prove to be a licensing nightmare for every AI developer, and in fact strongly privileges the tech giants over any start-up innovators, as only the former will have the means to engage in costly licensing on such large scale.¹¹² Also, one should be wary of the fact that the author will not necessarily benefit directly from this situation, as it will likely be the big rightholders that will license the uses, with the authors having to (re)negotiate successfully with their producers to get additional remuneration, which is not always an easy task.¹¹³

Therefore, it might be interesting to reflect on a possible statutory remuneration to the benefit of the author for the use of his works in the context of TDM activities for generative-AI purposes.¹¹⁴ In previous writings, we had proposed the replacement of the opt-out of Art. 4 CDSM by a statutory remuneration for commercial TDM activities,¹¹⁵ in order not to penalize start-ups developing useful AI systems in the EU.¹¹⁶ It needs to be recalled that AI systems cannot function without text and data mining, and therefore if we want to incentivize AI activities in the EU, TDM exceptions are crucial. This certainly does not mean that all the uses should be free in all circumstances, and in the EU there is vast experience with the “permitted-but-paid”¹¹⁷ model of remunerated limitations and exceptions (or, to use another terminology, “limitation-based remuneration rights”).¹¹⁸

Thus, by analogy with the idea of a commercial TDM activity, a specific remuneration right to the direct benefit of creators could be elaborated for the use of their work to train machines,¹¹⁹ possibly subjecting this right to mandatory collective management to make sure it can be rapidly implemented (in the context of the exception for private copy for example, large sums are collected and redistributed by collective management organizations to creators for the use of their works for private purposes, via a relatively well-functioning levy system¹²⁰). In this context, it would also make great sense to differentiate with regard to the purpose of the use and the works used in the machine learning process.¹²¹

Of course, the amount of the remuneration to be paid needs to be monitored closely (and preferably independently¹²²) in order not to create a significant hurdle for start-ups and AI developers to engage in this activity, and could be adjusted with regard to the economics of this sector and the potential losses which the original creator could incur (intuitively, but this would need to be verified by empirical studies, training AI on highly successful works on the market will create outputs that are likely to be commercially more successful, such as a song in “the style of” a renowned artist). Also, and this should not be forgotten, authors already use (and might increasingly use in the future) AI in their creative process as a tool for creativity, such as digital art using software or even more classically photography using cameras and filters. Opposing AI systems and authors systematically might not be a wise idea as they might very well cohabitate in the future and support each other.¹²³

The AI discussion might also be a good opportunity to reflect more generally on a specific remunerated exception for creative reuse that could englobe creative reuses of protected content, since in the digital environment (in particular, but it is also valid in the analog world) creators have increasingly incorporated protected elements in their creative process (sampling being a good example).¹²⁴ Of course, there might still be situations where the reuse of the work (by an AI or not) leads to unwanted results. If an AI is trained with protected works to issue outputs that are offensive, explicitly unwanted or carry inappropriate messages (racist, discriminatory, etc.), there should be room for authors to oppose it, but this could be dealt with – as already discussed supra – as a moral right issue and not prejudice any practical but fair solution under the economic rights.¹²⁵ The situation would not be radically different from a traditional moral rights violation, especially if the original work is recognizable.¹²⁶

7 Conclusion and Outlook

As we have seen, the development of AI generates a lot of questions.¹²⁷ How profoundly it will affect our lives – and in particularly creativity and the cultural ecosystems – remains to be seen. However, this is not the first – radical – technological (r)evolution that mankind has faced. The legal system, and in particularly the copyright system, (more or less) has always managed to adapt.¹²⁸ In any case, as in the past, the underlying human rights framework gives guidance and offers a workable compass in navigating reforms of the current copyright system with regard to generative AI systems.¹²⁹ What is certain is that AI invites a much deeper reflection than the question of the copyright issues of AI-generated works; and while lawyers are certainly concerned (as law frames the society we live in), possible future solutions should mostly be taken according to economic, philosophical, technological, artistic and ethical considerations.¹³⁰ Certainly, we are only at the beginning of an evolution. For now, as we have seen in this contribution, the copyrightability of AI-generated outputs is to be considered with utmost care¹³¹ and, as follows from a fundamental/human rights perspective, only when AI is used as a technical tool for creators in their creation process – meaning when they can serve a human author. However, we also need to be careful that the development of AI systems is not inhibited, as it can have a multitude of beneficial aspects if it is appropriately regulated.¹³² Thus, we have proposed the replacement of the opt-out mechanism of Art. 4 CDSM by a TDM exception for creative purposes coupled with a statutory remuneration to the benefit of authors only, in line with a proposal tabled in the past concerning a statutory remuneration for creative uses.¹³³ As we have tried to demonstrate, this remunerated “right to train the AI” can equally find support in the human rights framework.¹³⁴ Of course, this proposal will need to be developed, discussed, closely monitored and evaluated from a multidisciplinary perspective. What is certain is that copyright law should secure a vibrant environment for culture and creativity in the future. This can be done by regulating wisely these new technological environments, but this also requires (finally) cherishing and putting the human author¹³⁵ at the center of the copyright system (and not only the copyright industries). Doing this, we might in the future be able to have AI systems that serve creators and creativity, and not the other way around.